1// RUN: mlir-opt %s \ 2// RUN: -transform-interpreter \ 3// RUN: -test-transform-dialect-erase-schedule \ 4// RUN: | FileCheck %s 5 6memref.global "private" @bufferLhsGlobal : memref<64x32xf32, #gpu.address_space<workgroup>> 7memref.global "private" @bufferRhsGlobal : memref<8x32xf32, #gpu.address_space<workgroup>> 8 9// CHECK-LABEL: func.func @main() 10func.func @main() { 11 %c1 = arith.constant 1 : index 12 %c128 = arith.constant 128 : index 13 14 %0 = gpu.wait async 15 %memref, %asyncToken = gpu.alloc async [%0] () : memref<64x32xf32> 16 %memref_1, %asyncToken_2 = gpu.alloc async [%0] () : memref<8x32xf32> 17 18 // CHECK: %[[M1:.*]] = memref.cast %{{.*}} : memref<64x32xf32> to memref<*xf32> 19 // CHECK: %[[c64:.*]] = arith.constant 64 : index 20 // CHECK: %[[c32:.*]] = arith.constant 32 : index 21 // CHECK: %[[D1:.*]] = nvgpu.tma.create.descriptor %[[M1]] box[%[[c64]], %[[c32]]] 22 // CHECK-SAME: : memref<*xf32> -> <tensor = memref<64x32xf32, #gpu.address_space<workgroup>>, swizzle = none, l2promo = none, oob = zero, interleave = none> 23 // CHECK: %[[cast_2:.*]] = memref.cast %memref_0 : memref<8x32xf32> to memref<*xf32> 24 // CHECK: %[[c8_2:.*]] = arith.constant 8 : index 25 // CHECK: %[[c32_2:.*]] = arith.constant 32 : index 26 // CHECK: %[[D2:.*]] = nvgpu.tma.create.descriptor %cast_2 box[%[[c8_2]], %[[c32_2]]] 27 // CHECK-SAME: : memref<*xf32> -> <tensor = memref<8x32xf32, #gpu.address_space<workgroup>>, swizzle = none, l2promo = none, oob = zero, interleave = none> 28 // CHECK: gpu.launch 29 gpu.launch blocks(%bx, %by, %bz) in (%grid_x = %c1, %grid_y = %c1, %grid_z = %c1) 30 threads(%tx, %ty, %tz) in (%block_x = %c128, %block_y = %c1, %block_z = %c1) { 31 // CHECK: %[[G1:.*]] = memref.get_global @bufferLhsGlobal : memref<64x32xf32, #gpu.address_space<workgroup>> 32 // CHECK: %[[G2:.*]] = memref.get_global @bufferRhsGlobal : memref<8x32xf32, #gpu.address_space<workgroup>> 33 %out = memref.get_global @bufferLhsGlobal : memref<64x32xf32, #gpu.address_space<workgroup>> 34 %out_1 = memref.get_global @bufferRhsGlobal : memref<8x32xf32, #gpu.address_space<workgroup>> 35 36 // CHECK: %[[B:.*]] = nvgpu.mbarrier.create -> <memorySpace = #gpu.address_space<workgroup> 37 // CHECK: nvgpu.mbarrier.init %[[B]][%{{.*}}], %{{.*}} : <memorySpace = #gpu.address_space<workgroup> 38 // CHECK: gpu.barrier 39 // 40 // CHECK: %[[c0:.*]] = arith.constant 0 : index 41 // CHECK: %[[TIDX:.*]] = gpu.thread_id x 42 // CHECK: %[[CMP:.*]] = arith.cmpi eq, %[[TIDX]], %[[c0]] : index 43 // 44 // CHECK: scf.if %[[CMP]] { 45 // 46 // CHECK: %[[c0_7:.*]] = arith.constant 0 : index 47 // CHECK: nvgpu.tma.async.load %[[D1]][%[[c0_7]], %[[c0_7]]], %[[B]][%{{.*}}] to %[[G1]] 48 // CHECK-SAME: : <tensor = memref<64x32xf32, #gpu.address_space<workgroup>>, 49 // CHECK-SAME: swizzle = none, l2promo = none, oob = zero, interleave = none>, <memorySpace = #gpu.address_space<workgroup> 50 // CHECK-SAME: -> memref<64x32xf32, #gpu.address_space<workgroup>> 51 // 52 // CHECK: %[[c0_8:.*]] = arith.constant 0 : index 53 // CHECK: nvgpu.tma.async.load %[[D2]][%[[c0_8]], %[[c0_8]]], %[[B]][%{{.*}}] to %[[G2]] 54 // CHECK-SAME: : <tensor = memref<8x32xf32, #gpu.address_space<workgroup>>, 55 // CHECK-SAME: swizzle = none, l2promo = none, oob = zero, interleave = none>, <memorySpace = #gpu.address_space<workgroup> 56 // CHECK-SAME: -> memref<8x32xf32, #gpu.address_space<workgroup>> 57 // 58 // CHECK: %[[c9216:.*]] = arith.constant 9216 : index 59 // CHECK: nvgpu.mbarrier.arrive.expect_tx %[[B]][%{{.*}}], %[[c9216]] : <memorySpace = #gpu.address_space<workgroup> 60 // CHECK: } else { 61 // CHECK: %[[c0_7:.*]] = arith.constant 0 : index 62 // CHECK: nvgpu.mbarrier.arrive.expect_tx %[[B]][%{{.*}}], %[[c0_7]] : <memorySpace = #gpu.address_space<workgroup> 63 // CHECK: } 64 // 65 // CHECK: %[[c0_6:.*]] = llvm.mlir.constant(false) : i1 66 // CHECK: %[[c10000000:.*]] = arith.constant 10000000 : index 67 // CHECK: nvgpu.mbarrier.try_wait.parity %[[B]][%{{.*}}], %[[c0_6]], %[[c10000000]] : <memorySpace = #gpu.address_space<workgroup> 68 69 /// Both copies are matched and end up in the same async group. 70 linalg.copy ins(%memref: memref<64x32xf32>) outs(%out: memref<64x32xf32, #gpu.address_space<workgroup>>) 71 linalg.copy ins(%memref_1: memref<8x32xf32>) outs(%out_1: memref<8x32xf32, #gpu.address_space<workgroup>>) 72 73 gpu.terminator 74 } 75 76 return 77} 78 79module attributes {transform.with_named_sequence} { 80 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { 81 %copy = transform.structured.match ops{["linalg.copy"]} in %arg1 82 : (!transform.any_op) -> !transform.any_op 83 transform.nvgpu.rewrite_copy_as_tma %copy : (!transform.any_op) -> () 84 transform.yield 85 } 86} 87